US10824870B2 - Natural language eminence based robotic agent control - Google Patents
Natural language eminence based robotic agent control Download PDFInfo
- Publication number
- US10824870B2 US10824870B2 US16/020,611 US201816020611A US10824870B2 US 10824870 B2 US10824870 B2 US 10824870B2 US 201816020611 A US201816020611 A US 201816020611A US 10824870 B2 US10824870 B2 US 10824870B2
- Authority
- US
- United States
- Prior art keywords
- score
- insight
- eminence
- insights
- determining
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active, expires
Links
- 238000000034 method Methods 0.000 claims description 38
- 230000006870 function Effects 0.000 claims description 12
- 230000001419 dependent effect Effects 0.000 claims description 6
- 239000003795 chemical substances by application Substances 0.000 description 57
- 238000010586 diagram Methods 0.000 description 14
- 235000021178 picnic Nutrition 0.000 description 6
- 239000013598 vector Substances 0.000 description 6
- 238000005303 weighing Methods 0.000 description 5
- 230000001771 impaired effect Effects 0.000 description 4
- 239000000470 constituent Substances 0.000 description 3
- 230000001537 neural effect Effects 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 230000003716 rejuvenation Effects 0.000 description 2
- 239000011435 rock Substances 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- ABMMIOIRQJNRHG-XKNYDFJKSA-N Asn-Asn-Pro-Ser Chemical compound NC(=O)C[C@H](N)C(=O)N[C@@H](CC(N)=O)C(=O)N1CCC[C@H]1C(=O)N[C@@H](CO)C(O)=O ABMMIOIRQJNRHG-XKNYDFJKSA-N 0.000 description 1
- 241001503987 Clematis vitalba Species 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009194 climbing Effects 0.000 description 1
- 230000009193 crawling Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 235000013305 food Nutrition 0.000 description 1
- 230000014509 gene expression Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 230000002040 relaxant effect Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/35—Categorising the entire scene, e.g. birthday party or wedding scene
-
- G06K9/00684—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/903—Querying
- G06F16/9032—Query formulation
- G06F16/90332—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/216—Parsing using statistical methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/40—Processing or translation of natural language
- G06F40/55—Rule-based translation
- G06F40/56—Natural language generation
-
- G06K9/00671—
-
- G06K9/726—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/008—Artificial life, i.e. computing arrangements simulating life based on physical entities controlled by simulated intelligence so as to replicate intelligent life forms, e.g. based on robots replicating pets or humans in their appearance or behaviour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
- G06N5/022—Knowledge engineering; Knowledge acquisition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/20—Scenes; Scene-specific elements in augmented reality scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/26—Techniques for post-processing, e.g. correcting the recognition result
- G06V30/262—Techniques for post-processing, e.g. correcting the recognition result using context analysis, e.g. lexical, syntactic or semantic context
- G06V30/274—Syntactic or semantic context, e.g. balancing
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
- B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
- B25J9/00—Programme-controlled manipulators
- B25J9/16—Programme controls
- B25J9/1694—Programme controls characterised by use of sensors other than normal servo-feedback from position, speed or acceleration sensors, perception control, multi-sensor controlled systems, sensor fusion
- B25J9/1697—Vision controlled systems
-
- G06K9/00664—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
- G06N5/048—Fuzzy inferencing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
Definitions
- the robotic system may be pre-programmed with a set of instructions to perform a specified task, and/or to control a secondary device.
- the robotic system may obtain an image of an object or environment using a camera or another viewing device, and determine and/or receive, based on the image, a set of instructions.
- FIG. 1 illustrates a layout of a natural language eminence based robotic agent control apparatus in accordance with an example of the present disclosure
- FIG. 3 illustrates an image and possible insights in accordance with an example of the present disclosure
- FIG. 4 illustrates another image and possible insights to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 5 illustrates the aspect of multiple insights to multi-level insights to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 6 illustrates determination of the most reliable insight by determination of a reliability score to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 7 illustrates determination of the most atypical insight by determination of a degree of atypicalness to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 8 illustrates determination of the most concise insight by determination of a conciseness score to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 9 illustrates determination of an extent to which insights are to-the-point by determination of intrinsic and relative succinctness to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 10 illustrates determination of how natural insights are (e.g., determination of potentially inconsistent or erroneous insights) by transformation of each insight into a semantic graph and estimation of a naturalness score to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 11 illustrates intrinsic semantic relatedness graphs to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 12 illustrates estimation of semantic relatedness to identify duplicate (e.g., redundant) insights to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 13 illustrates further details of estimation of semantic relatedness to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 14 illustrates determination of central concepts embedded in insights to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 15 illustrates determination of semantic relatedness between insights to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 17 illustrates a dependency tree to illustrate operation of the natural language eminence based robotic agent control apparatus of FIG. 1 in accordance with an example of the present disclosure
- FIG. 18 illustrates an example block diagram for natural language eminence based robotic agent control in accordance with an example of the present disclosure
- FIG. 19 illustrates a flowchart of an example method for natural language eminence based robotic agent control in accordance with an example of the present disclosure.
- FIG. 20 illustrates a further example block diagram for natural language eminence based robotic agent control in accordance with another example of the present disclosure.
- the terms “a” and “an” are intended to denote at least one of a particular element.
- the term “includes” means includes but not limited to, the term “including” means including but not limited to.
- the term “based on” means based at least in part on.
- Natural language eminence based robotic agent control apparatuses, methods for natural language eminence based robotic agent control, and non-transitory computer readable media having stored thereon machine readable instructions to provide natural language eminence based robotic agent control are disclosed herein.
- the apparatuses, methods, and non-transitory computer readable media disclosed herein provide for natural language eminence based robotic agent control by ascertaining, by a robotic agent, an image of an object or an environment, and ascertaining a plurality of natural language insights for the image.
- a robotic agent may include a hardware and software device that may not be explicitly programmed to make decisions in uncertain environments (thus, it may be unknown what decisions a robotic agent may take in a new environment).
- an eminence score may be generated, and each insight of the plurality of insights may be ranked according to the eminence scores.
- An operation associated with the robotic agent, the object, or the environment may be controlled by the robotic agent and based on a highest ranked insight.
- an eminence score may be used to rank insights to make selections, for an absolute eminence score based analysis to perform computations on the eminence scores to make decisions (e.g., accept only those insights having a naturalness score >0.3), and/or a variability analysis on a set of eminence scores to perform inferences (e.g., complexity of the underlying object of observation).
- the apparatuses, methods, and non-transitory computer readable media disclosed herein provide for analysis of natural language text data (e.g., in the form of descriptions), for example, for images for visually impaired users, robotics, etc.
- natural language text data may be referred to as insights as disclosed herein.
- a robotic agent such as a machine, a vehicle, or another such device
- the robotic agent may obtain an image of an object or environment using a camera or another viewing device, and determine and/or receive, based on the image, a set of instructions.
- the instructions may be presented in the form of insights with respect to the image.
- a plurality of instructions may be presented in the form of insights to control the robotic agent, and/or to utilize the robotic agent to control a further device.
- the image may be analyzed to identify objects within the image.
- An image may also be analyzed to determine and/or ascertain insights with respect to the image and the identified objects.
- a plurality of insights are presented to the robotic agent with or without respect to an image, it is technically challenging for the robotic agent to eliminate uncertainties with respect to the plurality of insights, and to make a decision with respect to the plurality of insights, and/or with respect to the object or the environment being viewed by the robotic agent.
- the decision as disclosed herein may include performing a specified task such as manipulation of an object in the image, controlling a secondary device to perform a further task, and generally performing any type of operation that may be performed by a robotic agent.
- the apparatuses, methods, and non-transitory computer readable media disclosed herein may analyze an ensemble of multiple services to rank a multitude of outputs from different services.
- the apparatuses, methods, and non-transitory computer readable media disclosed herein may generate rankings of a plurality of insights for an end user and/or a robotic system from multiple heterogeneous insights that may be received and/or generated by different artificial intelligence services.
- the apparatuses, methods, and non-transitory computer readable media disclosed herein may present a ranking of a plurality of input insights to a user, such as a visually impaired user, by selecting the best description.
- the apparatuses, methods, and non-transitory computer readable media disclosed herein may present a ranking of a plurality of input insights (e.g., instructions) to control a robotic system and/or to be utilized by the robotic system to control a secondary device.
- the elements of the apparatuses, methods, and non-transitory computer readable media disclosed herein may be any combination of hardware and programming to implement the functionalities of the respective elements.
- the combinations of hardware and programming may be implemented in a number of different ways.
- the programming for the elements may be processor executable instructions stored on a non-transitory machine-readable storage medium and the hardware for the elements may include a processing resource to execute those instructions.
- a computing device implementing such elements may include the machine-readable storage medium storing the instructions and the processing resource to execute the instructions, or the machine-readable storage medium may be separately stored and accessible by the computing device and the processing resource.
- some elements may be implemented in circuitry.
- FIG. 1 illustrates a layout of an example natural language eminence based robotic agent control apparatus (hereinafter also referred to as “apparatus 100 ”).
- the apparatus 100 may include an insight analyzer 102 that is executed by at least one hardware processor (e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20 ) to ascertain, by a robotic agent 104 , an image 106 of an object 108 or an environment 110 , and ascertain a plurality of natural language insights 112 for the image 106 .
- at least one hardware processor e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20
- the apparatus 100 may include an insight analyzer 102 that is executed by at least one hardware processor (e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20 ) to ascertain, by a robotic agent 104 , an image 106 of an object 108 or an environment 110 , and ascertain a plurality of natural language insights 112 for the image 106 .
- An eminence score generator 114 that is executed by the at least one hardware processor (e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20 ) may generate, for each insight of the plurality of insights 112 , an eminence score 116 , and rank each insight of the plurality of insights 112 according to the eminence scores.
- the eminence score generator 114 may generate, for each insight of the plurality of insights 112 , the eminence score 116 by determining, for the eminence score 116 , a reliability score 118 by determining, by a semantic relatedness analyzer 120 that is executed by the at least one hardware processor (e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20 ), semantic relatedness 122 between each insight of the plurality of insights 112 .
- Semantic relatedness may indicate, for example, that two elements are related to each other, and semantic similarity may represent a specific type of semantic relatedness, which may indicate that the elements are equivalent as far as their usage is concerned.
- the semantic relatedness analyzer 120 may generate, based on the semantic relatedness 122 between each insight of the plurality of insights 112 , a semantic relatedness graph 124 , where each node of the semantic relatedness graph 124 may represent an insight of the plurality of insights 112 .
- the eminence score generator 114 may determine, for each node of the semantic relatedness graph 124 , a degree of centrality that represents the reliability score for the corresponding insight.
- An eminence score analyzer 126 that is executed by the at least one hardware processor (e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20 ) may analyze reliability scores for the plurality of insights 112 to identify at least one reliability score that exceeds a reliability score threshold. Further, the eminence score analyzer 126 may identify, for determination of a highest ranked insight 128 , at least one insight associated with the identified at least one reliability score that exceeds the reliability score threshold.
- the eminence score generator 114 may generate, for each insight of the plurality of insights 112 , the eminence score 116 by determining, for the eminence score 116 , a degree of atypicalness 130 by determining, for each insight of the plurality of insights 112 , by the semantic relatedness analyzer semantic relatedness between each pair of words of the insight, and determining, for each insight of the plurality of insights 112 , the degree of atypicalness 130 as a function of the semantic relatedness between each pair of words of the insight.
- the eminence score analyzer 126 may analyze degrees of atypicalness for the plurality of insights 112 to identify at least one degree of atypicalness that exceeds a degree of atypicalness threshold, and identify, for determination of the highest ranked insight, at least one insight associated with the identified at least one degree of atypicalness that exceeds the degree of atypicalness threshold.
- the eminence score generator 114 may generate, for each insight of the plurality of insights 112 , the eminence score 116 by determining, for the eminence score 116 , a conciseness score 132 by generating a concept graph that includes nodes that represent concepts extracted from the plurality of insights 112 , and edge weights that represent semantic relatedness between the concepts.
- the eminence score generator 114 may retain, for the concept graph, edges that include an edge weight that exceeds a specified edge weight threshold, generating groups based on remaining concepts that are connected by edges, and determining, for a specified insight, the conciseness score 132 as a function of a total number of concepts occurring in the specified insight and a total number of the groups that are spanned by the concepts occurring in the specified insight.
- the eminence score analyzer 126 may analyze conciseness scores for the plurality of insights 112 to identify at least one conciseness score that exceeds a conciseness score threshold, and identify, for determination of the highest ranked insight, at least one insight associated with the identified at least one conciseness score that exceeds the conciseness score threshold.
- the eminence score generator 114 may generate, for each insight of the plurality of insights 112 , the eminence score 116 by determining, for the eminence score 116 , an intrinsic succinctness score 134 by determining, for each insight of the plurality of insights 112 , noun type words.
- the eminence score generator 114 may generate, for each insight of the plurality of insights 112 , a dependency tree, determine, for each dependency tree, a number of dependent nodes associated with the noun type words, and determine, for each insight of the plurality of insights 112 , the intrinsic succinctness score 134 as a function of a number of the noun type words and the number of dependent nodes for the associated insight.
- the eminence score generator 114 may generate, for each insight of the plurality of insights 112 , the eminence score 116 by determining, for the eminence score 116 , a relative succinctness score 136 by determining, for each insight of the plurality of insights 112 , a hierarchy of concepts included in the insight, and determining a number of concepts included in a first insight of the plurality of insight that are at a higher level than concepts included in a second insight of the plurality of insights 112 .
- the eminence score analyzer 126 may analyze relative succinctness scores for the plurality of insights 112 to identify at least one relative succinctness score that exceeds a relative succinctness score threshold, and identify, for determination of the highest ranked insight, at least one insight associated with the identified at least one relative succinctness score that exceeds the relative succinctness score threshold.
- the eminence score generator 114 may generate, for each insight of the plurality of insights 112 , the eminence score 116 by determining, for the eminence score 116 , a naturalness score 138 by determining, for each insight of the plurality of insights 112 , a semantic relatedness between each pair of words in the insight, and determining, for each insight of the plurality of insights 112 , an expected semantic relatedness between node pairs in a semantic relatedness graph as an average of semantic relatedness scores across pairs of nodes in the semantic relatedness graph.
- the eminence score analyzer 126 may analyze naturalness scores for the plurality of insights 112 to identify at least one naturalness score that is less than a naturalness score threshold, and identify, for determination of the highest ranked insight, at least one remaining insight that is not associated with the identified at least one naturalness score that is less than the naturalness score threshold.
- the eminence score analyzer 126 may analyze, for each insight of the plurality of insights, a variability associated with the eminence score 116 , and determine, based on the variability associated with the eminence score, a degree of complexity of the image 106 .
- the eminence score generator 114 may generate, for each insight of the plurality of insights 112 , the eminence score 116 by determining, for the eminence score 116 , the reliability score 118 , the degree of atypicalness 130 , the conciseness score 132 , the succinctness score (e.g., intrinsic succinctness score 134 or relative succinctness score 136 ), and/or the naturalness score 138 , and determining whether the eminence score 116 exceeds a specified eminence score threshold.
- the succinctness score e.g., intrinsic succinctness score 134 or relative succinctness score 136
- the eminence score generator 114 may utilize an insight associated with the eminence score that exceeds the specified eminence score threshold for selection of the highest ranked insight.
- a robotic agent controller 140 that is executed by the at least one hardware processor (e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20 ) may control, by the robotic agent 104 and based on a highest ranked insight, an operation 142 associated with the robotic agent 104 , the object 108 , or the environment 110 .
- the at least one hardware processor e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20
- FIG. 2 illustrates a high-level process flow with respect to the apparatus 100 in accordance with an example of the present disclosure.
- the image 106 and a plurality of the insights 112 may be received.
- inputs to the apparatus 100 may include a set of insights (i.e., textual descriptions) about the image 106 , the object 108 , and/or the environment 110 .
- An insight may include one or more sentences, and/or a list of words or phrases. Examples of insights may include captions for images or videos, item reviews or feedbacks, social media comments on same event/news, and other such information.
- the insights 112 may be represented by ⁇ .
- may denote a number of insights in ⁇ .
- syntactic duplicates among insights may be removed before further processing, for example, by using approximate (or fuzzy) string matching techniques.
- the insights 112 may include at least two insights as inputs. However, with respect to the naturalness score 138 , and the succinctness score, the insights 112 may include at least one insight as input.
- Outputs of the apparatus 100 may include the eminence score 116 . Further, an output of the apparatus 100 may include a control signal to control the operation 142 associated with the robotic agent 104 .
- the eminence score generator 114 may determine the conciseness score 132 for each insight of the plurality of insights 112 , and identify, as results, an insight of the plurality of insights 112 with a maximum conciseness score. In this regard, the eminence score generator 114 may determine how comprehensive yet brief insights are.
- the eminence score generator 114 may determine the degree of atypicalness 130 for each insight of the plurality of insights 112 , and identify, as results, an insight of the plurality of insights 112 with a maximum degree of atypicalness. In this regard, the eminence score generator 114 may determine the degree of atypicalness 130 for each concept and insight.
- the eminence score generator 114 may determine the reliability score 118 for each insight of the plurality of insights 112 , and identify, as results, an insight of the plurality of insights 112 with a maximum reliability score. In this regard, the eminence score generator 114 may identify the most reliable insight.
- the eminence score generator 114 may determine the succinctness score (e.g., the intrinsic succinctness score 134 or the relative succinctness score 136 ) for each insight of the plurality of insights 112 , and identify, as results, an insight of the plurality of insights 112 with a maximum succinctness score. In this regard, the eminence score generator 114 may determine how brief insights are.
- the succinctness score e.g., the intrinsic succinctness score 134 or the relative succinctness score 136
- the eminence score generator 114 may determine the naturalness score 138 for each insight of the plurality of insights 112 , and identify, as results, an insight of the plurality of insights 112 with a maximum naturalness score. In this regard, the eminence score generator 114 may determine potentially erroneous or inconsistent insights that include low naturalness scores.
- the semantic relatedness analyzer 120 may perform tokenization and stop word removal for the insights 112 .
- the semantic relatedness analyzer 120 may extract tokens (e.g., words) from the insights.
- the semantic relatedness analyzer 120 may perform stop word removal, where stop words may include those words which are to be ignored during analysis.
- stop words may include those words which are to be ignored during analysis.
- a dynamically configurable list of stop words may be generated, or a predefined list of stop words for the language of insights may be used.
- the semantic relatedness analyzer 120 may perform term normalization by replacing all equivalent terms with one representative term.
- term normalization may include language thesaurus based synonym identification and unification (e.g., WORDNET based ‘goal’ and ‘objective’), lemmatization using, for example, language lexicon and morphological analysis (e.g., ‘movement’ and ‘moving’ including the same lemma as ‘move’), and short-form and long-form unification (e.g., ‘IP’ and ‘Intellectual Property’).
- stemming may be utilized prior to lemmatization (e.g., ‘trees’ being stemmed to ‘tree’).
- the semantic relatedness analyzer 120 may perform concept extraction by identifying potential ‘entity-terms’ as ‘noun-phrases’ and ‘functionalities’ as ‘verb-phrases’ by applying, for example, POS-TAGGERTM and CHUNKERTM. For example, in the sentence “Some trees near to a group of people camping in tents”, the identifiable entity terms may include “group of people”, “tents”, and “some trees”, and the identifiable functionality may include “camping”.
- the semantic relatedness analyzer 120 may utilize phrase mining to extract useful phrases from insights.
- the semantic relatedness analyzer 120 may determine vector space representations for words, phrases, and sentences.
- the semantic relatedness analyzer 120 may use word embeddings trained on external text corpuses.
- an external text corpus may include GLOVE, which may include global vectors for word representation, and may be trained, for example, on WIKIPEDIA, TWITTER, GIGAWORD, etc.
- WORD2VEC which includes neural embeddings for word representations, and may be trained, for example, on GOOGLE news corpus.
- the semantic relatedness analyzer 120 may further perform term relevance scoring.
- the term relevance scoring may represent a term weighing scheme that measures relative relevance of a term with respect to all other terms appearing in the insights 112 .
- Examples of term weighing schemes may include term frequency-inverse document frequency (Tf-Idf), etc.
- the term weighing schemes may estimate information theoretic weighing for each word with respect to available corpus of insights.
- rel(w) may be specified as the weight for word/phrase w, and rel(w) may be specified as:
- a may represent an empirical constant (e.g., 10 ⁇ 3 ), and p(w) may represent the probability of occurrence of w.
- Word embedding for each word may be updated as follows: v ( w ) ⁇ rel( w )* v ( w ) Equation (2)
- the semantic relatedness analyzer 120 may further analyze embeddings for concepts and insights.
- the semantic relatedness analyzer 120 may generate term embedding as an average of embeddings of the constituent words as follows:
- the semantic relatedness analyzer 120 may populate a list of words in ⁇ i as words( ⁇ i ), and determine the embedding for ⁇ i as a mean vector of its constituent words as follows:
- may represent a number of words in ⁇ i .
- Equation (3) and Equation (4) instead of mean(.), other functions such as min, max, etc., may be used.
- the semantic relatedness analyzer 120 may further perform semantic relatedness estimation for words.
- the semantic relatedness analyzer 120 may specify SemRel WordNet (w 1 , w 2 ) be the semantic relatedness estimation based, for example, upon an ontology, such as WORDNET.
- the semantic relatedness analyzer 120 may apply, for example, Lin measure, which provides normalized scores (i.e., in the range [0,1]), to estimate SemRel WordNet (w 1 , w 2 )
- Lin measure which provides normalized scores (i.e., in the range [0,1])
- SemRel WordNet w 1 , w 2
- the semantic relatedness analyzer 120 may further perform semantic relatedness estimation for multi-word text elements.
- the semantic relatedness analyzer 120 may specify X 1 and X 2 to be multi-word text elements, phrases (e.g., concepts), and insights.
- the semantic relatedness analyzer 120 may estimate the semantic relatedness score as described above.
- the semantic relatedness score between X 1 and X 2 may be defined as follows:
- SemRel ⁇ ( X 1 , X 2 ) 1 M + N ⁇ ( ⁇ w 1 ⁇ x ⁇ ⁇ max w 2 ⁇ y ⁇ ( SemRel ⁇ ( w 1 ⁇ x , w 2 ⁇ y ) ) + ⁇ w 2 ⁇ y ⁇ ⁇ max w 1 ⁇ x ⁇ ( SemRel ⁇ ( w 2 ⁇ y , w 1 ⁇ x ) ) ) Equation ⁇ ⁇ ( 7 )
- M and N may be described as the number of words occurring in X 1 and X 2 , respectively.
- the semantic relatedness analyzer 120 may perform joining of all the tokens in “entity terms” and “functional terms” using special character such as “_”, and replace those in the input corpus. For example, “this XYZ non-interventional study report” may be converted to “this_XYZ_non_interventional_study_report”.
- the semantic relatedness analyzer 120 may generate term embeddings. For example, since the number of insights for an image may be relatively small in number (e.g., less than 10 3 sentences), the semantic relatedness analyzer 120 may use word embeddings trained on external text corpuses. Examples of external text corpuses may include Global Vectors for Word Representation (GloVe), which may be trained on WIKIPEDIATM, TWITTERTM, GIGAWORDTM, etc. Other examples of external text corpuses may include Word2Vec (Neural embeddings for word representations), which may be trained on GOOGLE NEWS CORPUSTM.
- GloVe Global Vectors for Word Representation
- Other examples of external text corpuses may include Word2Vec (Neural embeddings for word representations), which may be trained on GOOGLE NEWS CORPUSTM.
- the semantic relatedness analyzer 120 may perform relatedness estimation as follows, for example, with respect to different insights that include insight #1, insight #2, and insight #3:
- the rel( ) function may be specified as follows:
- the eminence score generator 114 may generate the eminence score 116 for each insight of the plurality of insights 112 .
- the eminence score generator 114 may generate an undirected weighted graph G ⁇ (i.e., a semantic relatedness graph) with nodes representing insights, and semantic relatedness scores being used as weights associated with edges.
- the eminence score generator 114 may specify node n I to represent insight I.
- the eminence score generator 114 may determine the node's centrality score (by applying a node centrality technique, such as degree centrality, which is the average of all edge weights from a node, for weighted networks.
- the eminence score generator 114 may interpret reliability scores, where individual reliability scores may indicate a degree to which an insight has information/concepts that are supported by other insights.
- a high variation across insights may indicate that the underlying object of discussion (e.g., image) is potentially complex and consists of many semantically weakly related (or less known) aspects. Further, a lower variation may imply that either the underlying object of discussion is relatively simple or is well known.
- atypicalness ⁇ ( w ) 1 - ( avgConceptRel ⁇ ( w ) * Nor ⁇ ( iif ⁇ ( w ) ) ) Equation ⁇ ⁇ ( 14 )
- avgConceptRel ⁇ ( w ) mean w ′ ⁇ w ⁇ words ⁇ ( ⁇ ) ⁇ ⁇ SemRel ⁇ ( w , w ′ ) ⁇ Equation ⁇ ⁇ ( 15 )
- iif ⁇ ( w )
- Nor ⁇ ( iif ⁇ ( w ) ) iif ⁇ ( w ) - min w ⁇ words ⁇ ( ⁇ ) ⁇ ⁇ ( iif ⁇ ( w ) ) max w ⁇ words ⁇ (
- atypicalness scores may indicate a degree to which an insight is odd-one-out in the insight set.
- the atypicalness score may capture to what extent a current insight contains concepts which are semantically weakly related with most other concepts across insights.
- insights which are represented in a unique way may include higher scores in an atypicalness scale.
- high variation across insights may indicate that the underlying object of discussion (e.g., image) is observed to be associated with different types of aspects.
- the eminence score generator 114 may utilize, as a component of the eminence score 116 , the conciseness score 132 . With respect to the conciseness score 132 , the eminence score generator 114 may estimate conciseness by measuring how complete yet brief an insight is.
- the eminence score generator 114 may generate a global concept graph G words( ⁇ ) for which nodes may represent concepts extracted from insights, and edge weights may represent semantic relatedness scores between concepts (as disclosed herein with respect to semantic relatedness estimation for words).
- the eminence score generator 114 may merge semantically equivalent nodes in G words( ⁇ ) by retaining only those edges in G words( ⁇ ) that include a weight greater than d (e.g., 0.85).
- the eminence score generator 114 may collect all the nodes which are part of the same connected component in one group, resulting in partition of a set of concepts into very related concepts that are brought together in the same group (X p may represent the list of these groups). Further, the eminence score generator 114 may specify r as the number of total groups resulting from this process (e.g., the count of total number of semantically unique concepts across all insights).
- the eminence score generator 114 may specify i c as the total number of concept occurrences in I (repetitions of concepts may be counted as many times as they occur in the insight).
- the eminence score generator 114 may specify i u as the total number of groups in X p , which are spanned by the concepts in I (e.g., to count unique concepts present in the insight I).
- the eminence score generator 114 may determine the conciseness score 132 for an insight/as follows:
- i u r may measure relative completeness
- i u i c may measure degree of brevity (i.e., lack of redundancy).
- conciseness scores may indicate the degree to which an insight can be considered relatively complete. Higher conciseness scores (e.g., closer to 1) may indicate that the insight has low semantic redundancy among its descriptions, and the insight describes most of the aspects of the underlying object of discussion as compared to other insights. Lower conciseness scores may indicate that either the insight has high redundancy in its descriptions, or the insight is missing many of the aspects of the underlying object of discussion which are described in some other insights.
- high variation across insights may provide an indication on the nature of the insight set and the underlying sources.
- lower variation across insights may imply that most of the insights are received from technically similarly effective sources with respect to the concepts which are associated with the underlying object of discussion.
- the eminence score generator 114 may utilize, as a component of the eminence score 116 , the naturalness score 138 . With respect to the naturalness score 138 , for each insight, the eminence score generator 114 may determine semantic relatedness between each pair of words appearing within the insight (e.g., as disclosed herein with respect to semantic relatedness estimation for words). The determination of semantic relatedness between each pair of words may be used to determine an intrinsic semantic relatedness graph (ISG) for each insight, where nodes may represent words, and semantic relatedness scores may represent edge weights.
- ISG intrinsic semantic relatedness graph
- the eminence score generator 114 may determine expected semantic relatedness (referred to as the intrinsic semantic consistency (ISC) score) between any random pair of nodes in the intrinsic semantic graph as an average of semantic relatedness scores across a pair of nodes in the intrinsic semantic relatedness graph. The eminence score generator 114 may then determine the likelihood score of all part-of-speech (POS) trigrams within each insight.
- ISC intrinsic semantic consistency
- POS tagging may result into “some/DT trees/NNS near/IN to/TO a/DT group/NN of/IN people/NNS camping/VBG in/IN tents/NNS”, where the POS trigrams are ⁇ DT, NNS, IN ⁇ , ⁇ NNS, IN, TO ⁇ , ⁇ IN, TO, DT ⁇ , . . . , ⁇ VBG, IN, NNS ⁇ .
- the likelihood score of a trigram may represent the probability of these POS tags occurring together in a given order based upon the evidence present in a generic language model, such as WIKIPEDIA.
- the eminence score generator 114 may determine the part-of-speech score for the insight as the mean likelihood score across all trigrams in the insight. Further, the eminence score generator 114 may determine the naturalness score as an average of intrinsic semantic consistency score and the part-of-speech score.
- the naturalness score may indicate the degree to which an insight consists of terms which are strongly semantically related with one another (e.g., as captured by word embeddings trained on global knowledge bases).
- the naturalness score may also indicate how people or other intelligent agents (familiar with similar objects) are going to describe the object under observation in the same way as the current insight describes the object.
- a higher score may indicate that the insight includes most of the semantically strongly relevant concepts and has low redundancy among concepts contained in the insight.
- a lower score may indicate that the insight is describing those aspects of the underlying object of discussion, which are not so well related.
- high variation across insights may provide an indication that the underlying object of discussion (e.g., image) consists of multiple aspects, some of which are related with one another at various levels, while others are not found to be so closely related.
- lower variation may imply that either the underlying object of discussion is associated with most of the aspects which are naturally known to be together, or most of the objects are unrelated to one another.
- the eminence score generator 114 may utilize, as a component of the eminence score 116 , the succinctness score (e.g., intrinsic succinctness score 134 or relative succinctness score 136 ). With respect to the succinctness score, succinctness may measure how much to-the-point insights are. In order to determine succinctness, the eminence score generator 114 may determine two inter-related sub measures. Intrinsic succinctness may measure the degree to which an insight contains terms with minimum necessary details. Relative succinct may measure the degree to which an insight describes concepts using terms at higher levels of abstractions when compared with other insights describing same concept.
- the succinctness score e.g., intrinsic succinctness score 134 or relative succinctness score 136 .
- succinctness may measure how much to-the-point insights are.
- the eminence score generator 114 may determine two inter-related sub measures. Intrin
- the eminence score generator 114 may determine an intrinsic succinctness score as follows.
- the eminence score generator 114 may collect noun type words (with part of speech tags as NN (noun, singular or mass), NNS (noun, plural), NNP (proper noun, singular), NNPS (proper noun, plural), etc.
- the eminence score generator 114 may count dependent nodes for these noun type words.
- the intrinsic succinctness score of insight I may be determined as follows:
- Equation (19) may imply that intrinsic succinctness of an insight is high if entity terms appearing in the insight contain less further information.
- insight- 2 may be determined to be more succinct than insight- 1 since insight- 1 has additional details (e.g., “red shirt and green shorts” and “colorful”) reducing its succinctness.
- the eminence score generator 114 may specify c 1 , c 2 as the concepts appearing in the insights.
- the eminence score generator 114 may normalize ⁇ (.) scores to the [0,1] range by applying a min-max procedure.
- succinctness scores With respect to interpretation of succinctness scores, a higher succinctness score on a succinctness scale may indicate that the insight describes concepts at relatively higher levels of abstraction as compared to other insights, while using the minimum necessary details with terms.
- succinctness scores a high variation across insights may indicate that there are concepts associated with the underlying object of observation, which are being described at varying levels of abstractions and that different amounts of details are being given for concepts in different insights. This may mean that underlying sources of insights have very different technical foundations (e.g., learning model, training data, etc.), which is leading to such variations.
- a lower variation may imply that the underlying object of observation is associated with concepts which have relatively standard ways to describe them, and that sources of insights are behaviorally equivalent as far as their capability to generate expressions to convey these concept is of concern.
- the individual scores that include the reliability score, the degree of atypicalness, the conciseness score, the naturalness score, and/or the succinctness score may be totaled.
- the individual scores that include the reliability score, the degree of atypicalness, the conciseness score, the naturalness score, and/or the succinctness score may be normalized with respect to each other, and then a total eminence score may be determined to rank a plurality of unified insights.
- FIGS. 16 and 17 illustrate dependency trees to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- FIG. 3 illustrates an image and possible insights in accordance with an example of the present disclosure.
- the image may include a variety of objects.
- a variety of insights 112 may be specified for the image 106 .
- the insights may include the following:
- the various insights include errors such as incorrectness, incompleteness, OCR errors, etc.
- FIG. 4 illustrates another image and possible insights to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- the image of FIG. 4 may include various insights 112 , listed as “Insight1” to “Insight11”.
- FIG. 5 illustrates the aspect of multiple insights to multi-level insights to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- an analysis may be made as to how reliable insights are, how atypical insights are, how concise (i.e., comprehensive yet brief) insights are, to what extent insights are to the point, how natural insights are, intrinsic semantic consistency, what are the central concepts embedded in the insights, and how are insights semantically related to one another.
- FIG. 6 illustrates determination of the most reliable insight by determination of a reliability score to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- insight 1 may be determined to be least reliable, and insight 6 may be determined to be most reliable.
- insight 1 may include “A campground with a woman”, and insight 6 may include “A scenic picnic spot with lots of trees around, mountains nearby and group of people sitting around their tents.”
- the insights 112 may then be ranked according to their reliability score.
- FIG. 7 illustrates determination of the most atypical insight by determination of a degree of atypicalness to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- insight 1 may be determined to be least atypical
- insight 4 may be determined to be most atypical.
- insight 1 may include “A campground with a woman”
- insight 4 may include “Some trees near to a group of people camping in tents.”
- the eminence score generator 114 may determine the degree of atypicalness 130 for concepts, and thus insights.
- FIG. 8 illustrates determination of the most concise insight by determination of a conciseness score to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- insight 1 may be determined to be least concise, and insight 10 may be determined to be most concise.
- insight 1 may include “A campground with a woman”
- insight 10 may include “It's a beautiful place for an outing nice mountain view so many trees good place to chill out with family and friends.”
- FIG. 9 illustrates determination of an extent to which insights are to-the-point by determination of intrinsic and relative succinctness to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- insight 11 may be determined to be least succinct, and insight 3 may be determined to be most succinct.
- insight 11 may include “It's an awesome place for outing, rejuvenating and having fun. The view is awesome and full of greenery.
- the eminence score generator 114 may determine intrinsic and relative succinctness of entities, and in turn for insights.
- intrinsic succinctness may represent brevity
- relative succinctness may represent abstraction level.
- FIG. 10 illustrates determination of how natural insights are (e.g., determination of potentially inconsistent or erroneous insights) by transformation of each insight into a semantic graph and estimation of a naturalness score to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- natural insights e.g., determination of potentially inconsistent or erroneous insights
- insight 1 may be determined to be least consistent (e.g., least natural), and insight 3 may be determined to be most consistent (e.g., most natural).
- insight 1 may include “A campground with a woman”
- insight 3 may include “Picnic tables and chairs with tents on a campsite.”
- the naturalness score 138 may represent a degree of consistency.
- each insight may be transformed into a semantic relatedness graph, and a naturalness score may be determined for each insight to select insights with very low naturalness scores.
- Insight 3 may be determined to include a naturalness score 138 of 35.4%
- insight 1 may be determined to include a naturalness score 138 of 6.1%.
- Insight 3 and insight 1 may be determined to include relative naturalness scores of 100% and 17%.
- insight 3 may be determined to be a most consistent insight
- insight 1 may be determined to be a potentially inconsistent or erroneous insight.
- the error threshold Err threshold ISC
- ISC error threshold
- the naturalness score 138 may also be referred to as an intrinsic semantic consistency score, which may be determined based on a determination of semantic relatedness for each term of an insight, and then determining an average for an insight.
- the naturalness score for an insight may be determined by transforming the insight into an intrinsic semantic relatedness graph (e.g., as disclosed herein with respect to FIG. 11 ), and determining an intrinsic semantic consistency score.
- the likelihood score of all part-of-speech trigrams may be determined within the insight, and the mean likelihood score across all trigrams for the insight may represent a part-of-speech score for the insight.
- the naturalness score may be determined as an average of the intrinsic semantic consistency score and the part-of-speech score.
- insights may be ranked with respect to the eminence score, which may include the naturalness score.
- FIG. 11 illustrates intrinsic semantic relatedness graphs to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- the least consistent insight of FIG. 10 may include a semantic graph at 1100
- the most consistent insight of FIG. 10 may include a semantic graph at 1102 .
- the semantic relatedness graphs may also include the degree of semantic relatedness.
- FIG. 12 illustrates estimation of semantic relatedness to identify duplicate (e.g., redundant) insights to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- the semantic relatedness analyzer 120 may determine semantic relatedness among insights for eliminating redundancies.
- the semantic relatedness analyzer 120 may estimate semantic relatedness between each pair of insights to determine similar, or duplicate, or redundant insights.
- insight 5 may be determined to be redundant (e.g., subsumed by insight 7 ).
- representations may include neural embeddings, and upper ontology.
- measures may include Cosine, WordMover, JCN, etc.
- FIG. 13 illustrates further details of estimation of semantic relatedness to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- a semantic relatedness graph may be generated as shown at 1300 .
- the semantic relatedness graph may be used to estimate semantic relatedness.
- insight 5 A group of people camping together
- insight 7 A group of people camping and relaxing in the vicinity of nature
- insight 10 It's a beautiful place for an outing nice mountain view so many
- insight 11 It's an awesome place for outing, rejuvenating and having fun. The view is awesome and full of greenery. Perfect place for family and friends get together).
- insight 6 A scenic picnic spot with lots of trees around, mountains nearby and group of people sitting around their tents
- insight 4 Some trees near to a group of people camping in tents
- insight 9 Pane or outing involving camping. The ground is surrounded by a variety of green trees. There must be 20 people in the picnic
- insight 8 Some camping going on and food, chairs, tables and tents are around).
- FIG. 14 illustrates determination of central concepts embedded in insights to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- a central concepts identifier that is executed by the at least one hardware processor (e.g., the hardware processor 1802 of FIG. 18 , and/or the hardware processor 2004 of FIG. 20 ) may identify central concepts embedded within insights.
- the central concepts identifier may extract concepts from insights.
- the semantic relatedness analyzer 120 may determine semantic relatedness among insights for eliminating redundancies. For example, the semantic relatedness analyzer 120 may determine different semantic relatedness among concept terms across insights including semantic relatedness, ontological relations, and dependencies. Further, the semantic relatedness analyzer 120 may generate a concept graph for each insight using semantic relatedness among concept terms.
- the semantic relatedness analyzer 120 may generate the concept graph and determine (e.g., estimate) a semantic centrality measure for each concept term to determine central concepts.
- the central concepts identifier may identify central concepts embedded within insights as camp, people, tents, and trees.
- FIG. 15 illustrates determination of semantic relatedness between insights to illustrate operation of the apparatus 100 in accordance with an example of the present disclosure.
- FIGS. 18-20 respectively illustrate an example block diagram 1800 , a flowchart of an example method 1900 , and a further example block diagram 2000 for natural language eminence based robotic agent control, according to examples.
- the block diagram 1800 , the method 1900 , and the block diagram 2000 may be implemented on the apparatus 100 described above with reference to FIG. 1 by way of example and not of limitation.
- the block diagram 1800 , the method 1900 , and the block diagram 2000 may be practiced in other apparatus.
- FIG. 18 shows hardware of the apparatus 100 that may execute the instructions of the block diagram 1800 .
- the hardware may include a processor 1802 , and a memory 1804 storing machine readable instructions that when executed by the processor cause the processor to perform the instructions of the block diagram 1800 .
- the memory 1804 may represent a non-transitory computer readable medium.
- FIG. 19 may represent an example method for natural language eminence based robotic agent control, and the steps of the method.
- FIG. 20 may represent a non-transitory computer readable medium 2002 having stored thereon machine readable instructions to provide natural language eminence based robotic agent control according to an example. The machine readable instructions, when executed, cause a processor 2004 to perform the instructions of the block diagram 2000 also shown in FIG. 20 .
- the processor 1802 of FIG. 18 and/or the processor 2004 of FIG. 20 may include a single or multiple processors or other hardware processing circuit, to execute the methods, functions and other processes described herein. These methods, functions and other processes may be embodied as machine readable instructions stored on a computer readable medium, which may be non-transitory (e.g., the non-transitory computer readable medium 2002 of FIG. 20 ), such as hardware storage devices (e.g., RAM (random access memory), ROM (read only memory), EPROM (erasable, programmable ROM), EEPROM (electrically erasable, programmable ROM), hard drives, and flash memory).
- the memory 1804 may include a RAM, where the machine readable instructions and data for a processor may reside during runtime.
- the memory 1804 may include instructions 1806 to ascertain, by a robotic agent 104 , an image 106 of an object 108 or an environment 110 .
- the processor 1802 may fetch, decode, and execute the instructions 1808 to ascertain a plurality of natural language insights 112 for the image 106 .
- the processor 1802 may fetch, decode, and execute the instructions 1810 to generate, for each insight of the plurality of insights 112 , an eminence score 116 .
- the processor 1802 may fetch, decode, and execute the instructions 1812 to rank each insight of the plurality of insights 112 according to the eminence scores.
- the processor 1802 may fetch, decode, and execute the instructions 1808 to control, by the robotic agent 104 and based on a highest ranked insight 128 , an operation 142 associated with the robotic agent 104 , the object 108 , or the environment 110 .
- the method may include ascertaining, by a robotic agent 104 , an image 106 of an object 108 or an environment 110 .
- the method may include ascertaining, by at least one hardware processor, a plurality of natural language insights 112 for the image 106 .
- the method may include generating, by the at least one hardware processor, for each insight of the plurality of insights 112 , an eminence score 116 by determining, for the eminence score 116 , at least one of a reliability score, a degree of atypicalness, a conciseness score, a succinctness score, or a naturalness score, determining whether the eminence score 116 exceeds a specified eminence score 116 threshold, and based on a determination that the eminence score 116 exceeds the specified eminence for threshold, utilizing an insight associated with the eminence score 116 that exceeds the specified eminence score 116 threshold for selection of a highest ranked insight 128 .
- the method may include ranking, by the at least one hardware processor, each insight of the plurality of insights 112 according to the eminence scores.
- the method may include controlling, by the at least one hardware processor, by the robotic agent 104 and based on the highest ranked insight 128 , an operation 142 associated with the robotic agent 104 , the object 108 , or the environment 110 .
- the non-transitory computer readable medium 2002 may include instructions 2006 to ascertain a plurality of natural language insights 112 for an image 106 .
- the processor 2004 may fetch, decode, and execute the instructions 2008 to generate, for each insight of the plurality of insights 112 , an eminence score 116 .
- the processor 2004 may fetch, decode, and execute the instructions 2010 to rank each insight of the plurality of insights 112 according to the eminence scores.
- the processor 2004 may fetch, decode, and execute the instructions 2012 to control, by a robotic agent 104 and based on a highest ranked insight 128 , an operation 142 associated with the robotic agent 104 .
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Multimedia (AREA)
- Robotics (AREA)
- Probability & Statistics with Applications (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Animal Behavior & Ethology (AREA)
- Image Analysis (AREA)
- Machine Translation (AREA)
Abstract
Description
For Equation (1), a may represent an empirical constant (e.g., 10−3), and p(w) may represent the probability of occurrence of w. Word embedding for each word may be updated as follows:
v(w)←rel(w)*v(w) Equation (2)
For each insight ƒi∈Δ, the
For Equation (4), |words(ƒi)| may represent a number of words in ƒi. With respect to Equation (3) and Equation (4), instead of mean(.), other functions such as min, max, etc., may be used.
Relvec(w 1 ,w 2)=Cosine(v(w 1),v(w 2)) Equation (5)
The
SemRel(w 1 ,w 2)=max{Relvec(w 1 ,w 2),SemRelWordNet(w 1 ,w 2)} Equation (6)
v(w)←w bm25 *v(w) Equation (8)
For each multi-word term z=w1 . . . wn, the
v(z)←Σi=1 i=n v(w i) Equation (9)
v(ƒ i)=[v(entity(ƒ i)),v(action(ƒ i)),v(wd(ƒ i))] Equation (10)
For Equation (10):
-
- v(entity(ƒi)=Σz∈entity(ƒ
i )v(z) - v(action(ƒi)=Σz∈action(ƒ
i )v(z) - v(wd(ƒi))=Σw∈wd(ƒ
i )v(w)
- v(entity(ƒi)=Σz∈entity(ƒ
|
|
Insight #3 | ||
v(bdApp |
v(ddApp |
rel (v(bdApp |
||
-
- For two triplets of embedding vectors [X1e, X1a, X1w], [X2e, X2a, X2w] rel([X1e, X1a, X1w], [X2e, X2a, X2w])=[m(X1e, X2e), m(X1a, X2a), m(X1w, X2w)] m(.,.)=max{Cosine(.,.), WordMover(.,.)}
For each (I i ,I j≠i)∈Δ×Δ:w ij=SemRel(I i ,I j) Equation (11)
The
for each (w i ,w j≠i)∈words(Δ)×words(Δ):
δij=SemRel(w i ,w j) Equation(12)
For Equation (12), ε∈[0,1] may represent a constant for setting a lower threshold o n atypicalness of words (where a default may be set to 0.5).
atypicalness(I)=Σw∈I{atypicalness(w)>ε} Equation (13)
For Equation (13),
The
For Equation (18),
may measure relative completeness, and
may measure degree of brevity (i.e., lack of redundancy).
Equation (19) may imply that intrinsic succinctness of an insight is high if entity terms appearing in the insight contain less further information. For example, if insight-1 indicates that “A boy in red shirt and green shorts is playing with colorful ball,” and insight-2 indicates that “A boy is playing with a ball,” insight-2 may be determined to be more succinct than insight-1 since insight-1 has additional details (e.g., “red shirt and green shorts” and “colorful”) reducing its succinctness.
AbsLevelDiff(I 1 ,I 2)=Σc
For each insight I∈Δ, the
The
succinctness(I)=α*ISS(I)+(1−α)*RSS(I);α∈[0,1] Equation (22)
For Equation (22), α may represent a numeric parameter that may be configured externally in the range of 0 and 1, with a default value being specified, for example, as 0.5.
1. Person holding a sign | [Incorrect] | ||
2. Fast road sign | [Incomplete] | ||
3. Rizza | [OCR Errors] | ||
4. Bun | [Incomplete] | ||
Claims (20)
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
IN201741022771 | 2017-06-29 | ||
IN201741022771 | 2017-06-29 |
Publications (2)
Publication Number | Publication Date |
---|---|
US20190005329A1 US20190005329A1 (en) | 2019-01-03 |
US10824870B2 true US10824870B2 (en) | 2020-11-03 |
Family
ID=64738270
Family Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US16/020,566 Active 2039-05-26 US11062142B2 (en) | 2017-06-29 | 2018-06-27 | Natural language unification based robotic agent control |
US16/020,611 Active 2039-04-24 US10824870B2 (en) | 2017-06-29 | 2018-06-27 | Natural language eminence based robotic agent control |
Family Applications Before (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US16/020,566 Active 2039-05-26 US11062142B2 (en) | 2017-06-29 | 2018-06-27 | Natural language unification based robotic agent control |
Country Status (1)
Country | Link |
---|---|
US (2) | US11062142B2 (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10949765B2 (en) * | 2016-09-15 | 2021-03-16 | Accenture Global Solutions Limited | Automated inference of evidence from log information |
Families Citing this family (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10803248B1 (en) * | 2018-01-04 | 2020-10-13 | Facebook, Inc. | Consumer insights analysis using word embeddings |
US11314938B2 (en) * | 2019-07-29 | 2022-04-26 | Accenture Global Solutions Limited | Extracting actionable items from documents and assigning the actionable items to responsible parties |
CN110480635B (en) * | 2019-08-09 | 2020-12-04 | 中国人民解放军国防科技大学 | Control method and control system for multiple robots |
CN111582352B (en) * | 2020-04-30 | 2023-06-27 | 上海高仙自动化科技发展有限公司 | Object-based perception method, object-based perception device, robot and storage medium |
US11263407B1 (en) * | 2020-09-01 | 2022-03-01 | Rammer Technologies, Inc. | Determining topics and action items from conversations |
US11093718B1 (en) * | 2020-12-01 | 2021-08-17 | Rammer Technologies, Inc. | Determining conversational structure from speech |
US11594054B2 (en) * | 2021-02-19 | 2023-02-28 | Capital One Services, Llc | Document lineage management system |
US11302314B1 (en) | 2021-11-10 | 2022-04-12 | Rammer Technologies, Inc. | Tracking specialized concepts, topics, and activities in conversations |
US11599713B1 (en) | 2022-07-26 | 2023-03-07 | Rammer Technologies, Inc. | Summarizing conversational speech |
CN116229277B (en) * | 2023-05-08 | 2023-08-08 | 中国海洋大学 | Strong anti-interference ocean remote sensing image semantic segmentation method based on semantic correlation |
Citations (30)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5899956A (en) * | 1998-03-31 | 1999-05-04 | Advanced Future Technologies, Inc. | Vehicle mounted navigation device |
US20010021882A1 (en) * | 1999-12-31 | 2001-09-13 | Naoyasu Hosonuma | Robot apparatus and control method thereof |
US20040071347A1 (en) * | 2002-10-08 | 2004-04-15 | Frederick Kaplan | Adaptive artificial vision method and system |
US20040093328A1 (en) | 2001-02-08 | 2004-05-13 | Aditya Damle | Methods and systems for automated semantic knowledge leveraging graph theoretic analysis and the inherent structure of communication |
US20060143017A1 (en) * | 2004-12-24 | 2006-06-29 | Kabushiki Kaisha Toshiba | Interactive robot, speech recognition method and computer program product |
US20060222238A1 (en) * | 2005-03-31 | 2006-10-05 | Manabu Nishiyama | Image processing apparatus and image processing method |
US20070047809A1 (en) * | 2005-08-24 | 2007-03-01 | Denso Corporation | Environment recognition device |
US20070192910A1 (en) * | 2005-09-30 | 2007-08-16 | Clara Vu | Companion robot for personal interaction |
US20080119959A1 (en) * | 2006-11-21 | 2008-05-22 | Park Cheonshu | Expression of emotions in robot |
US20090232387A1 (en) * | 2007-01-03 | 2009-09-17 | Sandeep Gulati | Multi parallax exploitation for omni-directional imaging electronic eye |
US20100329504A1 (en) * | 2009-06-24 | 2010-12-30 | Xin Chen | Detecting Geographic Features in Images Based on Invariant Components |
US20110122247A1 (en) * | 2009-11-25 | 2011-05-26 | Lg Electronics Inc. | Surveillance system |
US8180760B1 (en) | 2007-12-20 | 2012-05-15 | Google Inc. | Organization system for ad campaigns |
US8200204B2 (en) * | 2008-07-28 | 2012-06-12 | Samsung Electronics Co., Ltd | Apparatus and method for displaying traffic information using widgets in mobile terminal |
US20120195499A1 (en) * | 2009-10-16 | 2012-08-02 | Nec Corporation | Color description analysis device, color description analysis method, and color description analysis program |
US20120242473A1 (en) * | 2011-03-25 | 2012-09-27 | Choi Sungha | Image processing for image dislay apparatus mounted to vehicle |
US20140157299A1 (en) * | 2012-11-30 | 2014-06-05 | Set Media, Inc. | Systems and Methods for Video-Level Reporting |
US20140280952A1 (en) | 2013-03-15 | 2014-09-18 | Advanced Elemental Technologies | Purposeful computing |
US9037296B2 (en) * | 2011-09-07 | 2015-05-19 | Lg Electronics Inc. | Robot cleaner, and system and method for remotely controlling the same |
US20150356414A1 (en) * | 2014-06-09 | 2015-12-10 | Cognitive Scale, Inc. | Universal Knowledge Repository |
US20160284217A1 (en) * | 2015-03-24 | 2016-09-29 | Lg Electronics Inc. | Vehicle, mobile terminal and method for controlling the same |
US20160378861A1 (en) | 2012-09-28 | 2016-12-29 | Sri International | Real-time human-machine collaboration using big data driven augmented reality technologies |
US20170040019A1 (en) | 2014-06-27 | 2017-02-09 | Kabushiki Kaisha Toshiba | Interaction apparatus and method |
US20170089714A1 (en) * | 2015-09-29 | 2017-03-30 | Xiaomi Inc. | Navigation method and device |
US20170116187A1 (en) * | 2015-10-22 | 2017-04-27 | International Business Machines Corporation | Natural language processor for providing natural language signals in a natural language output |
US20170125008A1 (en) * | 2014-04-17 | 2017-05-04 | Softbank Robotics Europe | Methods and systems of handling a dialog with a robot |
US20170361468A1 (en) * | 2016-06-15 | 2017-12-21 | Irobot Corporation | Systems and methods to control an autonomous mobile robot |
US20180012590A1 (en) * | 2016-07-08 | 2018-01-11 | Lg Electronics Inc. | Terminal and controlling method thereof |
US20180285386A1 (en) * | 2017-03-31 | 2018-10-04 | Alibaba Group Holding Limited | Method, apparatus, and electronic devices for searching images |
US20190158443A1 (en) * | 2017-11-17 | 2019-05-23 | International Business Machines Corporation | Real-time recommendation of message recipients based on recipient interest level in message |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2003065180A2 (en) | 2002-02-01 | 2003-08-07 | John Fairweather | System and method for creating a distributed network architecture |
US20160321259A1 (en) * | 2015-04-30 | 2016-11-03 | Linkedln Corporation | Network insights |
-
2018
- 2018-06-27 US US16/020,566 patent/US11062142B2/en active Active
- 2018-06-27 US US16/020,611 patent/US10824870B2/en active Active
Patent Citations (31)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5899956A (en) * | 1998-03-31 | 1999-05-04 | Advanced Future Technologies, Inc. | Vehicle mounted navigation device |
US20010021882A1 (en) * | 1999-12-31 | 2001-09-13 | Naoyasu Hosonuma | Robot apparatus and control method thereof |
US6542788B2 (en) * | 1999-12-31 | 2003-04-01 | Sony Corporation | Robot apparatus capable of selecting transmission destination, and control method therefor |
US20040093328A1 (en) | 2001-02-08 | 2004-05-13 | Aditya Damle | Methods and systems for automated semantic knowledge leveraging graph theoretic analysis and the inherent structure of communication |
US20040071347A1 (en) * | 2002-10-08 | 2004-04-15 | Frederick Kaplan | Adaptive artificial vision method and system |
US20060143017A1 (en) * | 2004-12-24 | 2006-06-29 | Kabushiki Kaisha Toshiba | Interactive robot, speech recognition method and computer program product |
US20060222238A1 (en) * | 2005-03-31 | 2006-10-05 | Manabu Nishiyama | Image processing apparatus and image processing method |
US20070047809A1 (en) * | 2005-08-24 | 2007-03-01 | Denso Corporation | Environment recognition device |
US20070192910A1 (en) * | 2005-09-30 | 2007-08-16 | Clara Vu | Companion robot for personal interaction |
US20080119959A1 (en) * | 2006-11-21 | 2008-05-22 | Park Cheonshu | Expression of emotions in robot |
US20090232387A1 (en) * | 2007-01-03 | 2009-09-17 | Sandeep Gulati | Multi parallax exploitation for omni-directional imaging electronic eye |
US8180760B1 (en) | 2007-12-20 | 2012-05-15 | Google Inc. | Organization system for ad campaigns |
US8200204B2 (en) * | 2008-07-28 | 2012-06-12 | Samsung Electronics Co., Ltd | Apparatus and method for displaying traffic information using widgets in mobile terminal |
US20100329504A1 (en) * | 2009-06-24 | 2010-12-30 | Xin Chen | Detecting Geographic Features in Images Based on Invariant Components |
US20120195499A1 (en) * | 2009-10-16 | 2012-08-02 | Nec Corporation | Color description analysis device, color description analysis method, and color description analysis program |
US20110122247A1 (en) * | 2009-11-25 | 2011-05-26 | Lg Electronics Inc. | Surveillance system |
US20120242473A1 (en) * | 2011-03-25 | 2012-09-27 | Choi Sungha | Image processing for image dislay apparatus mounted to vehicle |
US9037296B2 (en) * | 2011-09-07 | 2015-05-19 | Lg Electronics Inc. | Robot cleaner, and system and method for remotely controlling the same |
US20160378861A1 (en) | 2012-09-28 | 2016-12-29 | Sri International | Real-time human-machine collaboration using big data driven augmented reality technologies |
US20140157299A1 (en) * | 2012-11-30 | 2014-06-05 | Set Media, Inc. | Systems and Methods for Video-Level Reporting |
US20140280952A1 (en) | 2013-03-15 | 2014-09-18 | Advanced Elemental Technologies | Purposeful computing |
US20170125008A1 (en) * | 2014-04-17 | 2017-05-04 | Softbank Robotics Europe | Methods and systems of handling a dialog with a robot |
US20150356414A1 (en) * | 2014-06-09 | 2015-12-10 | Cognitive Scale, Inc. | Universal Knowledge Repository |
US20170040019A1 (en) | 2014-06-27 | 2017-02-09 | Kabushiki Kaisha Toshiba | Interaction apparatus and method |
US20160284217A1 (en) * | 2015-03-24 | 2016-09-29 | Lg Electronics Inc. | Vehicle, mobile terminal and method for controlling the same |
US20170089714A1 (en) * | 2015-09-29 | 2017-03-30 | Xiaomi Inc. | Navigation method and device |
US20170116187A1 (en) * | 2015-10-22 | 2017-04-27 | International Business Machines Corporation | Natural language processor for providing natural language signals in a natural language output |
US20170361468A1 (en) * | 2016-06-15 | 2017-12-21 | Irobot Corporation | Systems and methods to control an autonomous mobile robot |
US20180012590A1 (en) * | 2016-07-08 | 2018-01-11 | Lg Electronics Inc. | Terminal and controlling method thereof |
US20180285386A1 (en) * | 2017-03-31 | 2018-10-04 | Alibaba Group Holding Limited | Method, apparatus, and electronic devices for searching images |
US20190158443A1 (en) * | 2017-11-17 | 2019-05-23 | International Business Machines Corporation | Real-time recommendation of message recipients based on recipient interest level in message |
Non-Patent Citations (5)
Title |
---|
An overview of empirical natural language processing, Eric Brill et al., AI magazine, Pro Quest, Winter 1997, pp. 13-24 (Year: 1997). * |
Robot navigation based-Graphs, Christian Mandel et al., IEEE, 1-4244-0259, 2006, pp. 205-210 (Year: 2006). * |
Robot navigation based—Graphs, Christian Mandel et al., IEEE, 1-4244-0259, 2006, pp. 205-210 (Year: 2006). * |
Towards a construction-based-Simulation, Peter Ford Dominey, ELSEVIER, 0925-2312, 2006, pp. 2288-2302 (Year: 2006). * |
Towards a construction-based—Simulation, Peter Ford Dominey, ELSEVIER, 0925-2312, 2006, pp. 2288-2302 (Year: 2006). * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10949765B2 (en) * | 2016-09-15 | 2021-03-16 | Accenture Global Solutions Limited | Automated inference of evidence from log information |
Also Published As
Publication number | Publication date |
---|---|
US20190005328A1 (en) | 2019-01-03 |
US11062142B2 (en) | 2021-07-13 |
US20190005329A1 (en) | 2019-01-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10824870B2 (en) | Natural language eminence based robotic agent control | |
Zhang et al. | Ordinal common-sense inference | |
Schnabel et al. | Evaluation methods for unsupervised word embeddings | |
Shwartz et al. | Hypernyms under siege: Linguistically-motivated artillery for hypernymy detection | |
US11003638B2 (en) | System and method for building an evolving ontology from user-generated content | |
Fang et al. | From captions to visual concepts and back | |
CN112183994B (en) | Evaluation method and device for equipment state, computer equipment and storage medium | |
Firmanto et al. | Prediction of movie sentiment based on reviews and score on rotten tomatoes using sentiwordnet | |
Kaneko et al. | Dictionary-based debiasing of pre-trained word embeddings | |
Banjade et al. | Lemon and tea are not similar: Measuring word-to-word similarity by combining different methods | |
Cheema et al. | Check_square at checkthat! 2020: Claim detection in social media via fusion of transformer and syntactic features | |
Nabil et al. | Labr: A large scale arabic sentiment analysis benchmark | |
Friedman et al. | Finding dataset shortcuts with grammar induction | |
Singla et al. | Machine learning explanability method for the multi-label classification model | |
Nugraha et al. | Chatbot-Based Movie Recommender System Using POS Tagging | |
Vlachos et al. | Application-driven relation extraction with limited distant supervision | |
US20230075290A1 (en) | Method for linking a cve with at least one synthetic cpe | |
Leemann et al. | Coherence evaluation of visual concepts with objects and language | |
Soni et al. | Learning relational dependency networks for relation extraction | |
Wunnasri et al. | Solving unbalanced data for Thai sentiment analysis | |
Jiang et al. | Sentiment classification based on clause polarity and fusion via convolutional neural network | |
Mukuze et al. | A vision-grounded dataset for predicting typical locations for verbs | |
Prabhakar et al. | Question relevance in visual question answering | |
Tian et al. | A Random Walk Approach to Selectional Preferences Based on Preference Ranking and Propagation | |
Sahi et al. | Efficiency comparison of various plagiarism detection techniques |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
FEPP | Fee payment procedure |
Free format text: ENTITY STATUS SET TO UNDISCOUNTED (ORIGINAL EVENT CODE: BIG.); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY |
|
AS | Assignment |
Owner name: ACCENTURE GLOBAL SOLUTIONS LIMITED, IRELAND Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:MISRA, JANARDAN;PODDER, SANJAY;RAWAT, DIVYA;AND OTHERS;SIGNING DATES FROM 20180626 TO 20180820;REEL/FRAME:046940/0329 |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: DOCKETED NEW CASE - READY FOR EXAMINATION |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: NOTICE OF ALLOWANCE MAILED -- APPLICATION RECEIVED IN OFFICE OF PUBLICATIONS |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: PUBLICATIONS -- ISSUE FEE PAYMENT VERIFIED |
|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |
|
MAFP | Maintenance fee payment |
Free format text: PAYMENT OF MAINTENANCE FEE, 4TH YEAR, LARGE ENTITY (ORIGINAL EVENT CODE: M1551); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY Year of fee payment: 4 |